vm_pageout.c revision 208504
1/*-
2 * Copyright (c) 1991 Regents of the University of California.
3 * All rights reserved.
4 * Copyright (c) 1994 John S. Dyson
5 * All rights reserved.
6 * Copyright (c) 1994 David Greenman
7 * All rights reserved.
8 * Copyright (c) 2005 Yahoo! Technologies Norway AS
9 * All rights reserved.
10 *
11 * This code is derived from software contributed to Berkeley by
12 * The Mach Operating System project at Carnegie-Mellon University.
13 *
14 * Redistribution and use in source and binary forms, with or without
15 * modification, are permitted provided that the following conditions
16 * are met:
17 * 1. Redistributions of source code must retain the above copyright
18 *    notice, this list of conditions and the following disclaimer.
19 * 2. Redistributions in binary form must reproduce the above copyright
20 *    notice, this list of conditions and the following disclaimer in the
21 *    documentation and/or other materials provided with the distribution.
22 * 3. All advertising materials mentioning features or use of this software
23 *    must display the following acknowledgement:
24 *	This product includes software developed by the University of
25 *	California, Berkeley and its contributors.
26 * 4. Neither the name of the University nor the names of its contributors
27 *    may be used to endorse or promote products derived from this software
28 *    without specific prior written permission.
29 *
30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33 * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40 * SUCH DAMAGE.
41 *
42 *	from: @(#)vm_pageout.c	7.4 (Berkeley) 5/7/91
43 *
44 *
45 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
46 * All rights reserved.
47 *
48 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
49 *
50 * Permission to use, copy, modify and distribute this software and
51 * its documentation is hereby granted, provided that both the copyright
52 * notice and this permission notice appear in all copies of the
53 * software, derivative works or modified versions, and any portions
54 * thereof, and that both notices appear in supporting documentation.
55 *
56 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
57 * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
58 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
59 *
60 * Carnegie Mellon requests users of this software to return to
61 *
62 *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
63 *  School of Computer Science
64 *  Carnegie Mellon University
65 *  Pittsburgh PA 15213-3890
66 *
67 * any improvements or extensions that they make and grant Carnegie the
68 * rights to redistribute these changes.
69 */
70
71/*
72 *	The proverbial page-out daemon.
73 */
74
75#include <sys/cdefs.h>
76__FBSDID("$FreeBSD: head/sys/vm/vm_pageout.c 208504 2010-05-24 14:26:57Z alc $");
77
78#include "opt_vm.h"
79#include <sys/param.h>
80#include <sys/systm.h>
81#include <sys/kernel.h>
82#include <sys/eventhandler.h>
83#include <sys/lock.h>
84#include <sys/mutex.h>
85#include <sys/proc.h>
86#include <sys/kthread.h>
87#include <sys/ktr.h>
88#include <sys/mount.h>
89#include <sys/resourcevar.h>
90#include <sys/sched.h>
91#include <sys/signalvar.h>
92#include <sys/vnode.h>
93#include <sys/vmmeter.h>
94#include <sys/sx.h>
95#include <sys/sysctl.h>
96
97#include <vm/vm.h>
98#include <vm/vm_param.h>
99#include <vm/vm_object.h>
100#include <vm/vm_page.h>
101#include <vm/vm_map.h>
102#include <vm/vm_pageout.h>
103#include <vm/vm_pager.h>
104#include <vm/swap_pager.h>
105#include <vm/vm_extern.h>
106#include <vm/uma.h>
107
108/*
109 * System initialization
110 */
111
112/* the kernel process "vm_pageout"*/
113static void vm_pageout(void);
114static int vm_pageout_clean(vm_page_t);
115static void vm_pageout_scan(int pass);
116
117struct proc *pageproc;
118
119static struct kproc_desc page_kp = {
120	"pagedaemon",
121	vm_pageout,
122	&pageproc
123};
124SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start,
125    &page_kp);
126
127#if !defined(NO_SWAPPING)
128/* the kernel process "vm_daemon"*/
129static void vm_daemon(void);
130static struct	proc *vmproc;
131
132static struct kproc_desc vm_kp = {
133	"vmdaemon",
134	vm_daemon,
135	&vmproc
136};
137SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
138#endif
139
140
141int vm_pages_needed;		/* Event on which pageout daemon sleeps */
142int vm_pageout_deficit;		/* Estimated number of pages deficit */
143int vm_pageout_pages_needed;	/* flag saying that the pageout daemon needs pages */
144
145#if !defined(NO_SWAPPING)
146static int vm_pageout_req_swapout;	/* XXX */
147static int vm_daemon_needed;
148static struct mtx vm_daemon_mtx;
149/* Allow for use by vm_pageout before vm_daemon is initialized. */
150MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
151#endif
152static int vm_max_launder = 32;
153static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
154static int vm_pageout_full_stats_interval = 0;
155static int vm_pageout_algorithm=0;
156static int defer_swap_pageouts=0;
157static int disable_swap_pageouts=0;
158
159#if defined(NO_SWAPPING)
160static int vm_swap_enabled=0;
161static int vm_swap_idle_enabled=0;
162#else
163static int vm_swap_enabled=1;
164static int vm_swap_idle_enabled=0;
165#endif
166
167SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
168	CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
169
170SYSCTL_INT(_vm, OID_AUTO, max_launder,
171	CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
172
173SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
174	CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
175
176SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
177	CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
178
179SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
180	CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
181
182#if defined(NO_SWAPPING)
183SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
184	CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout");
185SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
186	CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
187#else
188SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
189	CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
190SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
191	CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
192#endif
193
194SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
195	CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
196
197SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
198	CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
199
200static int pageout_lock_miss;
201SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
202	CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
203
204#define VM_PAGEOUT_PAGE_COUNT 16
205int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
206
207int vm_page_max_wired;		/* XXX max # of wired pages system-wide */
208SYSCTL_INT(_vm, OID_AUTO, max_wired,
209	CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count");
210
211#if !defined(NO_SWAPPING)
212static void vm_pageout_map_deactivate_pages(vm_map_t, long);
213static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
214static void vm_req_vmdaemon(int req);
215#endif
216static void vm_pageout_page_stats(void);
217
218static void
219vm_pageout_init_marker(vm_page_t marker, u_short queue)
220{
221
222	bzero(marker, sizeof(*marker));
223	marker->flags = PG_FICTITIOUS | PG_MARKER;
224	marker->oflags = VPO_BUSY;
225	marker->queue = queue;
226	marker->wire_count = 1;
227}
228
229/*
230 * vm_pageout_fallback_object_lock:
231 *
232 * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is
233 * known to have failed and page queue must be either PQ_ACTIVE or
234 * PQ_INACTIVE.  To avoid lock order violation, unlock the page queues
235 * while locking the vm object.  Use marker page to detect page queue
236 * changes and maintain notion of next page on page queue.  Return
237 * TRUE if no changes were detected, FALSE otherwise.  vm object is
238 * locked on return.
239 *
240 * This function depends on both the lock portion of struct vm_object
241 * and normal struct vm_page being type stable.
242 */
243boolean_t
244vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
245{
246	struct vm_page marker;
247	boolean_t unchanged;
248	u_short queue;
249	vm_object_t object;
250
251	queue = m->queue;
252	vm_pageout_init_marker(&marker, queue);
253	object = m->object;
254
255	TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl,
256			   m, &marker, pageq);
257	vm_page_unlock_queues();
258	vm_page_unlock(m);
259	VM_OBJECT_LOCK(object);
260	vm_page_lock(m);
261	vm_page_lock_queues();
262
263	/* Page queue might have changed. */
264	*next = TAILQ_NEXT(&marker, pageq);
265	unchanged = (m->queue == queue &&
266		     m->object == object &&
267		     &marker == TAILQ_NEXT(m, pageq));
268	TAILQ_REMOVE(&vm_page_queues[queue].pl,
269		     &marker, pageq);
270	return (unchanged);
271}
272
273/*
274 * Lock the page while holding the page queue lock.  Use marker page
275 * to detect page queue changes and maintain notion of next page on
276 * page queue.  Return TRUE if no changes were detected, FALSE
277 * otherwise.  The page is locked on return. The page queue lock might
278 * be dropped and reacquired.
279 *
280 * This function depends on normal struct vm_page being type stable.
281 */
282boolean_t
283vm_pageout_page_lock(vm_page_t m, vm_page_t *next)
284{
285	struct vm_page marker;
286	boolean_t unchanged;
287	u_short queue;
288
289	vm_page_lock_assert(m, MA_NOTOWNED);
290	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
291
292	if (vm_page_trylock(m))
293		return (TRUE);
294
295	queue = m->queue;
296	vm_pageout_init_marker(&marker, queue);
297
298	TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl, m, &marker, pageq);
299	vm_page_unlock_queues();
300	vm_page_lock(m);
301	vm_page_lock_queues();
302
303	/* Page queue might have changed. */
304	*next = TAILQ_NEXT(&marker, pageq);
305	unchanged = (m->queue == queue && &marker == TAILQ_NEXT(m, pageq));
306	TAILQ_REMOVE(&vm_page_queues[queue].pl, &marker, pageq);
307	return (unchanged);
308}
309
310/*
311 * vm_pageout_clean:
312 *
313 * Clean the page and remove it from the laundry.
314 *
315 * We set the busy bit to cause potential page faults on this page to
316 * block.  Note the careful timing, however, the busy bit isn't set till
317 * late and we cannot do anything that will mess with the page.
318 */
319static int
320vm_pageout_clean(vm_page_t m)
321{
322	vm_object_t object;
323	vm_page_t mc[2*vm_pageout_page_count];
324	int pageout_count;
325	int ib, is, page_base;
326	vm_pindex_t pindex = m->pindex;
327
328	vm_page_lock_assert(m, MA_NOTOWNED);
329	vm_page_lock(m);
330	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
331
332	/*
333	 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
334	 * with the new swapper, but we could have serious problems paging
335	 * out other object types if there is insufficient memory.
336	 *
337	 * Unfortunately, checking free memory here is far too late, so the
338	 * check has been moved up a procedural level.
339	 */
340
341	/*
342	 * Can't clean the page if it's busy or held.
343	 */
344	if ((m->hold_count != 0) ||
345	    ((m->busy != 0) || (m->oflags & VPO_BUSY))) {
346		vm_page_unlock(m);
347		return 0;
348	}
349
350	mc[vm_pageout_page_count] = m;
351	pageout_count = 1;
352	page_base = vm_pageout_page_count;
353	ib = 1;
354	is = 1;
355
356	/*
357	 * Scan object for clusterable pages.
358	 *
359	 * We can cluster ONLY if: ->> the page is NOT
360	 * clean, wired, busy, held, or mapped into a
361	 * buffer, and one of the following:
362	 * 1) The page is inactive, or a seldom used
363	 *    active page.
364	 * -or-
365	 * 2) we force the issue.
366	 *
367	 * During heavy mmap/modification loads the pageout
368	 * daemon can really fragment the underlying file
369	 * due to flushing pages out of order and not trying
370	 * align the clusters (which leave sporatic out-of-order
371	 * holes).  To solve this problem we do the reverse scan
372	 * first and attempt to align our cluster, then do a
373	 * forward scan if room remains.
374	 */
375	object = m->object;
376more:
377	while (ib && pageout_count < vm_pageout_page_count) {
378		vm_page_t p;
379
380		if (ib > pindex) {
381			ib = 0;
382			break;
383		}
384
385		if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
386			ib = 0;
387			break;
388		}
389		if ((p->oflags & VPO_BUSY) || p->busy) {
390			ib = 0;
391			break;
392		}
393		vm_page_lock(p);
394		vm_page_test_dirty(p);
395		if (p->dirty == 0 ||
396		    p->queue != PQ_INACTIVE ||
397		    p->hold_count != 0) {	/* may be undergoing I/O */
398			vm_page_unlock(p);
399			ib = 0;
400			break;
401		}
402		vm_page_unlock(p);
403		mc[--page_base] = p;
404		++pageout_count;
405		++ib;
406		/*
407		 * alignment boundry, stop here and switch directions.  Do
408		 * not clear ib.
409		 */
410		if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
411			break;
412	}
413
414	while (pageout_count < vm_pageout_page_count &&
415	    pindex + is < object->size) {
416		vm_page_t p;
417
418		if ((p = vm_page_lookup(object, pindex + is)) == NULL)
419			break;
420		if ((p->oflags & VPO_BUSY) || p->busy) {
421			break;
422		}
423		vm_page_lock(p);
424		vm_page_test_dirty(p);
425		if (p->dirty == 0 ||
426		    p->queue != PQ_INACTIVE ||
427		    p->hold_count != 0) {	/* may be undergoing I/O */
428			vm_page_unlock(p);
429			break;
430		}
431		vm_page_unlock(p);
432		mc[page_base + pageout_count] = p;
433		++pageout_count;
434		++is;
435	}
436
437	/*
438	 * If we exhausted our forward scan, continue with the reverse scan
439	 * when possible, even past a page boundry.  This catches boundry
440	 * conditions.
441	 */
442	if (ib && pageout_count < vm_pageout_page_count)
443		goto more;
444
445	vm_page_unlock(m);
446	/*
447	 * we allow reads during pageouts...
448	 */
449	return (vm_pageout_flush(&mc[page_base], pageout_count, 0));
450}
451
452/*
453 * vm_pageout_flush() - launder the given pages
454 *
455 *	The given pages are laundered.  Note that we setup for the start of
456 *	I/O ( i.e. busy the page ), mark it read-only, and bump the object
457 *	reference count all in here rather then in the parent.  If we want
458 *	the parent to do more sophisticated things we may have to change
459 *	the ordering.
460 */
461int
462vm_pageout_flush(vm_page_t *mc, int count, int flags)
463{
464	vm_object_t object = mc[0]->object;
465	int pageout_status[count];
466	int numpagedout = 0;
467	int i;
468
469	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
470	mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
471
472	/*
473	 * Initiate I/O.  Bump the vm_page_t->busy counter and
474	 * mark the pages read-only.
475	 *
476	 * We do not have to fixup the clean/dirty bits here... we can
477	 * allow the pager to do it after the I/O completes.
478	 *
479	 * NOTE! mc[i]->dirty may be partial or fragmented due to an
480	 * edge case with file fragments.
481	 */
482	for (i = 0; i < count; i++) {
483		KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
484		    ("vm_pageout_flush: partially invalid page %p index %d/%d",
485			mc[i], i, count));
486		vm_page_io_start(mc[i]);
487		pmap_remove_write(mc[i]);
488	}
489	vm_object_pip_add(object, count);
490
491	vm_pager_put_pages(object, mc, count, flags, pageout_status);
492
493	for (i = 0; i < count; i++) {
494		vm_page_t mt = mc[i];
495
496		KASSERT(pageout_status[i] == VM_PAGER_PEND ||
497		    (mt->flags & PG_WRITEABLE) == 0,
498		    ("vm_pageout_flush: page %p is not write protected", mt));
499		switch (pageout_status[i]) {
500		case VM_PAGER_OK:
501		case VM_PAGER_PEND:
502			numpagedout++;
503			break;
504		case VM_PAGER_BAD:
505			/*
506			 * Page outside of range of object. Right now we
507			 * essentially lose the changes by pretending it
508			 * worked.
509			 */
510			vm_page_undirty(mt);
511			break;
512		case VM_PAGER_ERROR:
513		case VM_PAGER_FAIL:
514			/*
515			 * If page couldn't be paged out, then reactivate the
516			 * page so it doesn't clog the inactive list.  (We
517			 * will try paging out it again later).
518			 */
519			vm_page_lock(mt);
520			vm_page_activate(mt);
521			vm_page_unlock(mt);
522			break;
523		case VM_PAGER_AGAIN:
524			break;
525		}
526
527		/*
528		 * If the operation is still going, leave the page busy to
529		 * block all other accesses. Also, leave the paging in
530		 * progress indicator set so that we don't attempt an object
531		 * collapse.
532		 */
533		if (pageout_status[i] != VM_PAGER_PEND) {
534			vm_object_pip_wakeup(object);
535			vm_page_io_finish(mt);
536			if (vm_page_count_severe()) {
537				vm_page_lock(mt);
538				vm_page_try_to_cache(mt);
539				vm_page_unlock(mt);
540			}
541		}
542	}
543	return (numpagedout);
544}
545
546#if !defined(NO_SWAPPING)
547/*
548 *	vm_pageout_object_deactivate_pages
549 *
550 *	deactivate enough pages to satisfy the inactive target
551 *	requirements or if vm_page_proc_limit is set, then
552 *	deactivate all of the pages in the object and its
553 *	backing_objects.
554 *
555 *	The object and map must be locked.
556 */
557static void
558vm_pageout_object_deactivate_pages(pmap, first_object, desired)
559	pmap_t pmap;
560	vm_object_t first_object;
561	long desired;
562{
563	vm_object_t backing_object, object;
564	vm_page_t p, next;
565	int actcount, remove_mode;
566
567	VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
568	if (first_object->type == OBJT_DEVICE ||
569	    first_object->type == OBJT_SG)
570		return;
571	for (object = first_object;; object = backing_object) {
572		if (pmap_resident_count(pmap) <= desired)
573			goto unlock_return;
574		VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
575		if (object->type == OBJT_PHYS || object->paging_in_progress)
576			goto unlock_return;
577
578		remove_mode = 0;
579		if (object->shadow_count > 1)
580			remove_mode = 1;
581		/*
582		 * scan the objects entire memory queue
583		 */
584		p = TAILQ_FIRST(&object->memq);
585		while (p != NULL) {
586			if (pmap_resident_count(pmap) <= desired)
587				goto unlock_return;
588			next = TAILQ_NEXT(p, listq);
589			if ((p->oflags & VPO_BUSY) != 0 || p->busy != 0) {
590				p = next;
591				continue;
592			}
593			vm_page_lock(p);
594			vm_page_lock_queues();
595			cnt.v_pdpages++;
596			if (p->wire_count != 0 ||
597			    p->hold_count != 0 ||
598			    !pmap_page_exists_quick(pmap, p)) {
599				vm_page_unlock_queues();
600				vm_page_unlock(p);
601				p = next;
602				continue;
603			}
604			actcount = pmap_ts_referenced(p);
605			if (actcount) {
606				vm_page_flag_set(p, PG_REFERENCED);
607			} else if (p->flags & PG_REFERENCED) {
608				actcount = 1;
609			}
610			if ((p->queue != PQ_ACTIVE) &&
611				(p->flags & PG_REFERENCED)) {
612				vm_page_activate(p);
613				p->act_count += actcount;
614				vm_page_flag_clear(p, PG_REFERENCED);
615			} else if (p->queue == PQ_ACTIVE) {
616				if ((p->flags & PG_REFERENCED) == 0) {
617					p->act_count -= min(p->act_count, ACT_DECLINE);
618					if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) {
619						pmap_remove_all(p);
620						vm_page_deactivate(p);
621					} else {
622						vm_page_requeue(p);
623					}
624				} else {
625					vm_page_activate(p);
626					vm_page_flag_clear(p, PG_REFERENCED);
627					if (p->act_count < (ACT_MAX - ACT_ADVANCE))
628						p->act_count += ACT_ADVANCE;
629					vm_page_requeue(p);
630				}
631			} else if (p->queue == PQ_INACTIVE) {
632				pmap_remove_all(p);
633			}
634			vm_page_unlock_queues();
635			vm_page_unlock(p);
636			p = next;
637		}
638		if ((backing_object = object->backing_object) == NULL)
639			goto unlock_return;
640		VM_OBJECT_LOCK(backing_object);
641		if (object != first_object)
642			VM_OBJECT_UNLOCK(object);
643	}
644unlock_return:
645	if (object != first_object)
646		VM_OBJECT_UNLOCK(object);
647}
648
649/*
650 * deactivate some number of pages in a map, try to do it fairly, but
651 * that is really hard to do.
652 */
653static void
654vm_pageout_map_deactivate_pages(map, desired)
655	vm_map_t map;
656	long desired;
657{
658	vm_map_entry_t tmpe;
659	vm_object_t obj, bigobj;
660	int nothingwired;
661
662	if (!vm_map_trylock(map))
663		return;
664
665	bigobj = NULL;
666	nothingwired = TRUE;
667
668	/*
669	 * first, search out the biggest object, and try to free pages from
670	 * that.
671	 */
672	tmpe = map->header.next;
673	while (tmpe != &map->header) {
674		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
675			obj = tmpe->object.vm_object;
676			if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) {
677				if (obj->shadow_count <= 1 &&
678				    (bigobj == NULL ||
679				     bigobj->resident_page_count < obj->resident_page_count)) {
680					if (bigobj != NULL)
681						VM_OBJECT_UNLOCK(bigobj);
682					bigobj = obj;
683				} else
684					VM_OBJECT_UNLOCK(obj);
685			}
686		}
687		if (tmpe->wired_count > 0)
688			nothingwired = FALSE;
689		tmpe = tmpe->next;
690	}
691
692	if (bigobj != NULL) {
693		vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
694		VM_OBJECT_UNLOCK(bigobj);
695	}
696	/*
697	 * Next, hunt around for other pages to deactivate.  We actually
698	 * do this search sort of wrong -- .text first is not the best idea.
699	 */
700	tmpe = map->header.next;
701	while (tmpe != &map->header) {
702		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
703			break;
704		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
705			obj = tmpe->object.vm_object;
706			if (obj != NULL) {
707				VM_OBJECT_LOCK(obj);
708				vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
709				VM_OBJECT_UNLOCK(obj);
710			}
711		}
712		tmpe = tmpe->next;
713	}
714
715	/*
716	 * Remove all mappings if a process is swapped out, this will free page
717	 * table pages.
718	 */
719	if (desired == 0 && nothingwired) {
720		pmap_remove(vm_map_pmap(map), vm_map_min(map),
721		    vm_map_max(map));
722	}
723	vm_map_unlock(map);
724}
725#endif		/* !defined(NO_SWAPPING) */
726
727/*
728 *	vm_pageout_scan does the dirty work for the pageout daemon.
729 */
730static void
731vm_pageout_scan(int pass)
732{
733	vm_page_t m, next;
734	struct vm_page marker;
735	int page_shortage, maxscan, pcount;
736	int addl_page_shortage, addl_page_shortage_init;
737	vm_object_t object;
738	int actcount;
739	int vnodes_skipped = 0;
740	int maxlaunder;
741
742	/*
743	 * Decrease registered cache sizes.
744	 */
745	EVENTHANDLER_INVOKE(vm_lowmem, 0);
746	/*
747	 * We do this explicitly after the caches have been drained above.
748	 */
749	uma_reclaim();
750
751	addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit);
752
753	/*
754	 * Calculate the number of pages we want to either free or move
755	 * to the cache.
756	 */
757	page_shortage = vm_paging_target() + addl_page_shortage_init;
758
759	vm_pageout_init_marker(&marker, PQ_INACTIVE);
760
761	/*
762	 * Start scanning the inactive queue for pages we can move to the
763	 * cache or free.  The scan will stop when the target is reached or
764	 * we have scanned the entire inactive queue.  Note that m->act_count
765	 * is not used to form decisions for the inactive queue, only for the
766	 * active queue.
767	 *
768	 * maxlaunder limits the number of dirty pages we flush per scan.
769	 * For most systems a smaller value (16 or 32) is more robust under
770	 * extreme memory and disk pressure because any unnecessary writes
771	 * to disk can result in extreme performance degredation.  However,
772	 * systems with excessive dirty pages (especially when MAP_NOSYNC is
773	 * used) will die horribly with limited laundering.  If the pageout
774	 * daemon cannot clean enough pages in the first pass, we let it go
775	 * all out in succeeding passes.
776	 */
777	if ((maxlaunder = vm_max_launder) <= 1)
778		maxlaunder = 1;
779	if (pass)
780		maxlaunder = 10000;
781	vm_page_lock_queues();
782rescan0:
783	addl_page_shortage = addl_page_shortage_init;
784	maxscan = cnt.v_inactive_count;
785
786	for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
787	     m != NULL && maxscan-- > 0 && page_shortage > 0;
788	     m = next) {
789
790		cnt.v_pdpages++;
791
792		if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE) {
793			goto rescan0;
794		}
795
796		next = TAILQ_NEXT(m, pageq);
797
798		/*
799		 * skip marker pages
800		 */
801		if (m->flags & PG_MARKER)
802			continue;
803
804		/*
805		 * Lock the page.
806		 */
807		if (!vm_pageout_page_lock(m, &next)) {
808			vm_page_unlock(m);
809			addl_page_shortage++;
810			continue;
811		}
812
813		/*
814		 * A held page may be undergoing I/O, so skip it.
815		 */
816		if (m->hold_count || (object = m->object) == NULL) {
817			vm_page_unlock(m);
818			vm_page_requeue(m);
819			addl_page_shortage++;
820			continue;
821		}
822
823		/*
824		 * Don't mess with busy pages, keep in the front of the
825		 * queue, most likely are being paged out.
826		 */
827		if (!VM_OBJECT_TRYLOCK(object) &&
828		    (!vm_pageout_fallback_object_lock(m, &next) ||
829			m->hold_count != 0)) {
830			VM_OBJECT_UNLOCK(object);
831			vm_page_unlock(m);
832			addl_page_shortage++;
833			continue;
834		}
835		if (m->busy || (m->oflags & VPO_BUSY)) {
836			vm_page_unlock(m);
837			VM_OBJECT_UNLOCK(object);
838			addl_page_shortage++;
839			continue;
840		}
841
842		/*
843		 * If the object is not being used, we ignore previous
844		 * references.
845		 */
846		if (object->ref_count == 0) {
847			vm_page_flag_clear(m, PG_REFERENCED);
848			KASSERT(!pmap_page_is_mapped(m),
849			    ("vm_pageout_scan: page %p is mapped", m));
850
851		/*
852		 * Otherwise, if the page has been referenced while in the
853		 * inactive queue, we bump the "activation count" upwards,
854		 * making it less likely that the page will be added back to
855		 * the inactive queue prematurely again.  Here we check the
856		 * page tables (or emulated bits, if any), given the upper
857		 * level VM system not knowing anything about existing
858		 * references.
859		 */
860		} else if (((m->flags & PG_REFERENCED) == 0) &&
861			(actcount = pmap_ts_referenced(m))) {
862			vm_page_activate(m);
863			VM_OBJECT_UNLOCK(object);
864			m->act_count += (actcount + ACT_ADVANCE);
865			vm_page_unlock(m);
866			continue;
867		}
868
869		/*
870		 * If the upper level VM system knows about any page
871		 * references, we activate the page.  We also set the
872		 * "activation count" higher than normal so that we will less
873		 * likely place pages back onto the inactive queue again.
874		 */
875		if ((m->flags & PG_REFERENCED) != 0) {
876			vm_page_flag_clear(m, PG_REFERENCED);
877			actcount = pmap_ts_referenced(m);
878			vm_page_activate(m);
879			VM_OBJECT_UNLOCK(object);
880			m->act_count += (actcount + ACT_ADVANCE + 1);
881			vm_page_unlock(m);
882			continue;
883		}
884
885		/*
886		 * If the upper level VM system does not believe that the page
887		 * is fully dirty, but it is mapped for write access, then we
888		 * consult the pmap to see if the page's dirty status should
889		 * be updated.
890		 */
891		if (m->dirty != VM_PAGE_BITS_ALL &&
892		    (m->flags & PG_WRITEABLE) != 0) {
893			/*
894			 * Avoid a race condition: Unless write access is
895			 * removed from the page, another processor could
896			 * modify it before all access is removed by the call
897			 * to vm_page_cache() below.  If vm_page_cache() finds
898			 * that the page has been modified when it removes all
899			 * access, it panics because it cannot cache dirty
900			 * pages.  In principle, we could eliminate just write
901			 * access here rather than all access.  In the expected
902			 * case, when there are no last instant modifications
903			 * to the page, removing all access will be cheaper
904			 * overall.
905			 */
906			if (pmap_is_modified(m))
907				vm_page_dirty(m);
908			else if (m->dirty == 0)
909				pmap_remove_all(m);
910		}
911
912		if (m->valid == 0) {
913			/*
914			 * Invalid pages can be easily freed
915			 */
916			vm_page_free(m);
917			cnt.v_dfree++;
918			--page_shortage;
919		} else if (m->dirty == 0) {
920			/*
921			 * Clean pages can be placed onto the cache queue.
922			 * This effectively frees them.
923			 */
924			vm_page_cache(m);
925			--page_shortage;
926		} else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
927			/*
928			 * Dirty pages need to be paged out, but flushing
929			 * a page is extremely expensive verses freeing
930			 * a clean page.  Rather then artificially limiting
931			 * the number of pages we can flush, we instead give
932			 * dirty pages extra priority on the inactive queue
933			 * by forcing them to be cycled through the queue
934			 * twice before being flushed, after which the
935			 * (now clean) page will cycle through once more
936			 * before being freed.  This significantly extends
937			 * the thrash point for a heavily loaded machine.
938			 */
939			vm_page_flag_set(m, PG_WINATCFLS);
940			vm_page_requeue(m);
941		} else if (maxlaunder > 0) {
942			/*
943			 * We always want to try to flush some dirty pages if
944			 * we encounter them, to keep the system stable.
945			 * Normally this number is small, but under extreme
946			 * pressure where there are insufficient clean pages
947			 * on the inactive queue, we may have to go all out.
948			 */
949			int swap_pageouts_ok, vfslocked = 0;
950			struct vnode *vp = NULL;
951			struct mount *mp = NULL;
952
953			if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
954				swap_pageouts_ok = 1;
955			} else {
956				swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
957				swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
958				vm_page_count_min());
959
960			}
961
962			/*
963			 * We don't bother paging objects that are "dead".
964			 * Those objects are in a "rundown" state.
965			 */
966			if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
967				vm_page_unlock(m);
968				VM_OBJECT_UNLOCK(object);
969				vm_page_requeue(m);
970				continue;
971			}
972
973			/*
974			 * Following operations may unlock
975			 * vm_page_queue_mtx, invalidating the 'next'
976			 * pointer.  To prevent an inordinate number
977			 * of restarts we use our marker to remember
978			 * our place.
979			 *
980			 */
981			TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl,
982					   m, &marker, pageq);
983			/*
984			 * The object is already known NOT to be dead.   It
985			 * is possible for the vget() to block the whole
986			 * pageout daemon, but the new low-memory handling
987			 * code should prevent it.
988			 *
989			 * The previous code skipped locked vnodes and, worse,
990			 * reordered pages in the queue.  This results in
991			 * completely non-deterministic operation and, on a
992			 * busy system, can lead to extremely non-optimal
993			 * pageouts.  For example, it can cause clean pages
994			 * to be freed and dirty pages to be moved to the end
995			 * of the queue.  Since dirty pages are also moved to
996			 * the end of the queue once-cleaned, this gives
997			 * way too large a weighting to defering the freeing
998			 * of dirty pages.
999			 *
1000			 * We can't wait forever for the vnode lock, we might
1001			 * deadlock due to a vn_read() getting stuck in
1002			 * vm_wait while holding this vnode.  We skip the
1003			 * vnode if we can't get it in a reasonable amount
1004			 * of time.
1005			 */
1006			if (object->type == OBJT_VNODE) {
1007				vm_page_unlock_queues();
1008				vm_page_unlock(m);
1009				vp = object->handle;
1010				if (vp->v_type == VREG &&
1011				    vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1012					mp = NULL;
1013					++pageout_lock_miss;
1014					if (object->flags & OBJ_MIGHTBEDIRTY)
1015						vnodes_skipped++;
1016					vm_page_lock_queues();
1017					goto unlock_and_continue;
1018				}
1019				KASSERT(mp != NULL,
1020				    ("vp %p with NULL v_mount", vp));
1021				vm_object_reference_locked(object);
1022				VM_OBJECT_UNLOCK(object);
1023				vfslocked = VFS_LOCK_GIANT(vp->v_mount);
1024				if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK,
1025				    curthread)) {
1026					VM_OBJECT_LOCK(object);
1027					vm_page_lock_queues();
1028					++pageout_lock_miss;
1029					if (object->flags & OBJ_MIGHTBEDIRTY)
1030						vnodes_skipped++;
1031					vp = NULL;
1032					goto unlock_and_continue;
1033				}
1034				VM_OBJECT_LOCK(object);
1035				vm_page_lock(m);
1036				vm_page_lock_queues();
1037				/*
1038				 * The page might have been moved to another
1039				 * queue during potential blocking in vget()
1040				 * above.  The page might have been freed and
1041				 * reused for another vnode.
1042				 */
1043				if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE ||
1044				    m->object != object ||
1045				    TAILQ_NEXT(m, pageq) != &marker) {
1046					vm_page_unlock(m);
1047					if (object->flags & OBJ_MIGHTBEDIRTY)
1048						vnodes_skipped++;
1049					goto unlock_and_continue;
1050				}
1051
1052				/*
1053				 * The page may have been busied during the
1054				 * blocking in vget().  We don't move the
1055				 * page back onto the end of the queue so that
1056				 * statistics are more correct if we don't.
1057				 */
1058				if (m->busy || (m->oflags & VPO_BUSY)) {
1059					vm_page_unlock(m);
1060					goto unlock_and_continue;
1061				}
1062
1063				/*
1064				 * If the page has become held it might
1065				 * be undergoing I/O, so skip it
1066				 */
1067				if (m->hold_count) {
1068					vm_page_unlock(m);
1069					vm_page_requeue(m);
1070					if (object->flags & OBJ_MIGHTBEDIRTY)
1071						vnodes_skipped++;
1072					goto unlock_and_continue;
1073				}
1074			}
1075			vm_page_unlock(m);
1076
1077			/*
1078			 * If a page is dirty, then it is either being washed
1079			 * (but not yet cleaned) or it is still in the
1080			 * laundry.  If it is still in the laundry, then we
1081			 * start the cleaning operation.
1082			 *
1083			 * decrement page_shortage on success to account for
1084			 * the (future) cleaned page.  Otherwise we could wind
1085			 * up laundering or cleaning too many pages.
1086			 */
1087			vm_page_unlock_queues();
1088			if (vm_pageout_clean(m) != 0) {
1089				--page_shortage;
1090				--maxlaunder;
1091			}
1092			vm_page_lock_queues();
1093unlock_and_continue:
1094			vm_page_lock_assert(m, MA_NOTOWNED);
1095			VM_OBJECT_UNLOCK(object);
1096			if (mp != NULL) {
1097				vm_page_unlock_queues();
1098				if (vp != NULL)
1099					vput(vp);
1100				VFS_UNLOCK_GIANT(vfslocked);
1101				vm_object_deallocate(object);
1102				vn_finished_write(mp);
1103				vm_page_lock_queues();
1104			}
1105			next = TAILQ_NEXT(&marker, pageq);
1106			TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl,
1107				     &marker, pageq);
1108			vm_page_lock_assert(m, MA_NOTOWNED);
1109			continue;
1110		}
1111		vm_page_unlock(m);
1112		VM_OBJECT_UNLOCK(object);
1113	}
1114
1115	/*
1116	 * Compute the number of pages we want to try to move from the
1117	 * active queue to the inactive queue.
1118	 */
1119	page_shortage = vm_paging_target() +
1120		cnt.v_inactive_target - cnt.v_inactive_count;
1121	page_shortage += addl_page_shortage;
1122
1123	/*
1124	 * Scan the active queue for things we can deactivate. We nominally
1125	 * track the per-page activity counter and use it to locate
1126	 * deactivation candidates.
1127	 */
1128	pcount = cnt.v_active_count;
1129	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1130	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1131
1132	while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1133
1134		KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE),
1135		    ("vm_pageout_scan: page %p isn't active", m));
1136
1137		next = TAILQ_NEXT(m, pageq);
1138		object = m->object;
1139		if ((m->flags & PG_MARKER) != 0) {
1140			m = next;
1141			continue;
1142		}
1143		if (!vm_pageout_page_lock(m, &next) ||
1144		    (object = m->object) == NULL) {
1145			vm_page_unlock(m);
1146			m = next;
1147			continue;
1148		}
1149		if (!VM_OBJECT_TRYLOCK(object) &&
1150		    !vm_pageout_fallback_object_lock(m, &next)) {
1151			VM_OBJECT_UNLOCK(object);
1152			vm_page_unlock(m);
1153			m = next;
1154			continue;
1155		}
1156
1157		/*
1158		 * Don't deactivate pages that are busy.
1159		 */
1160		if ((m->busy != 0) ||
1161		    (m->oflags & VPO_BUSY) ||
1162		    (m->hold_count != 0)) {
1163			vm_page_unlock(m);
1164			VM_OBJECT_UNLOCK(object);
1165			vm_page_requeue(m);
1166			m = next;
1167			continue;
1168		}
1169
1170		/*
1171		 * The count for pagedaemon pages is done after checking the
1172		 * page for eligibility...
1173		 */
1174		cnt.v_pdpages++;
1175
1176		/*
1177		 * Check to see "how much" the page has been used.
1178		 */
1179		actcount = 0;
1180		if (object->ref_count != 0) {
1181			if (m->flags & PG_REFERENCED) {
1182				actcount += 1;
1183			}
1184			actcount += pmap_ts_referenced(m);
1185			if (actcount) {
1186				m->act_count += ACT_ADVANCE + actcount;
1187				if (m->act_count > ACT_MAX)
1188					m->act_count = ACT_MAX;
1189			}
1190		}
1191
1192		/*
1193		 * Since we have "tested" this bit, we need to clear it now.
1194		 */
1195		vm_page_flag_clear(m, PG_REFERENCED);
1196
1197		/*
1198		 * Only if an object is currently being used, do we use the
1199		 * page activation count stats.
1200		 */
1201		if (actcount && (object->ref_count != 0)) {
1202			vm_page_requeue(m);
1203		} else {
1204			m->act_count -= min(m->act_count, ACT_DECLINE);
1205			if (vm_pageout_algorithm ||
1206			    object->ref_count == 0 ||
1207			    m->act_count == 0) {
1208				page_shortage--;
1209				if (object->ref_count == 0) {
1210					KASSERT(!pmap_page_is_mapped(m),
1211				    ("vm_pageout_scan: page %p is mapped", m));
1212					if (m->dirty == 0)
1213						vm_page_cache(m);
1214					else
1215						vm_page_deactivate(m);
1216				} else {
1217					vm_page_deactivate(m);
1218				}
1219			} else {
1220				vm_page_requeue(m);
1221			}
1222		}
1223		vm_page_unlock(m);
1224		VM_OBJECT_UNLOCK(object);
1225		m = next;
1226	}
1227	vm_page_unlock_queues();
1228#if !defined(NO_SWAPPING)
1229	/*
1230	 * Idle process swapout -- run once per second.
1231	 */
1232	if (vm_swap_idle_enabled) {
1233		static long lsec;
1234		if (time_second != lsec) {
1235			vm_req_vmdaemon(VM_SWAP_IDLE);
1236			lsec = time_second;
1237		}
1238	}
1239#endif
1240
1241	/*
1242	 * If we didn't get enough free pages, and we have skipped a vnode
1243	 * in a writeable object, wakeup the sync daemon.  And kick swapout
1244	 * if we did not get enough free pages.
1245	 */
1246	if (vm_paging_target() > 0) {
1247		if (vnodes_skipped && vm_page_count_min())
1248			(void) speedup_syncer();
1249#if !defined(NO_SWAPPING)
1250		if (vm_swap_enabled && vm_page_count_target())
1251			vm_req_vmdaemon(VM_SWAP_NORMAL);
1252#endif
1253	}
1254
1255	/*
1256	 * If we are critically low on one of RAM or swap and low on
1257	 * the other, kill the largest process.  However, we avoid
1258	 * doing this on the first pass in order to give ourselves a
1259	 * chance to flush out dirty vnode-backed pages and to allow
1260	 * active pages to be moved to the inactive queue and reclaimed.
1261	 */
1262	if (pass != 0 &&
1263	    ((swap_pager_avail < 64 && vm_page_count_min()) ||
1264	     (swap_pager_full && vm_paging_target() > 0)))
1265		vm_pageout_oom(VM_OOM_MEM);
1266}
1267
1268
1269void
1270vm_pageout_oom(int shortage)
1271{
1272	struct proc *p, *bigproc;
1273	vm_offset_t size, bigsize;
1274	struct thread *td;
1275	struct vmspace *vm;
1276
1277	/*
1278	 * We keep the process bigproc locked once we find it to keep anyone
1279	 * from messing with it; however, there is a possibility of
1280	 * deadlock if process B is bigproc and one of it's child processes
1281	 * attempts to propagate a signal to B while we are waiting for A's
1282	 * lock while walking this list.  To avoid this, we don't block on
1283	 * the process lock but just skip a process if it is already locked.
1284	 */
1285	bigproc = NULL;
1286	bigsize = 0;
1287	sx_slock(&allproc_lock);
1288	FOREACH_PROC_IN_SYSTEM(p) {
1289		int breakout;
1290
1291		if (PROC_TRYLOCK(p) == 0)
1292			continue;
1293		/*
1294		 * If this is a system, protected or killed process, skip it.
1295		 */
1296		if ((p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM)) ||
1297		    (p->p_pid == 1) || P_KILLED(p) ||
1298		    ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1299			PROC_UNLOCK(p);
1300			continue;
1301		}
1302		/*
1303		 * If the process is in a non-running type state,
1304		 * don't touch it.  Check all the threads individually.
1305		 */
1306		breakout = 0;
1307		FOREACH_THREAD_IN_PROC(p, td) {
1308			thread_lock(td);
1309			if (!TD_ON_RUNQ(td) &&
1310			    !TD_IS_RUNNING(td) &&
1311			    !TD_IS_SLEEPING(td)) {
1312				thread_unlock(td);
1313				breakout = 1;
1314				break;
1315			}
1316			thread_unlock(td);
1317		}
1318		if (breakout) {
1319			PROC_UNLOCK(p);
1320			continue;
1321		}
1322		/*
1323		 * get the process size
1324		 */
1325		vm = vmspace_acquire_ref(p);
1326		if (vm == NULL) {
1327			PROC_UNLOCK(p);
1328			continue;
1329		}
1330		if (!vm_map_trylock_read(&vm->vm_map)) {
1331			vmspace_free(vm);
1332			PROC_UNLOCK(p);
1333			continue;
1334		}
1335		size = vmspace_swap_count(vm);
1336		vm_map_unlock_read(&vm->vm_map);
1337		if (shortage == VM_OOM_MEM)
1338			size += vmspace_resident_count(vm);
1339		vmspace_free(vm);
1340		/*
1341		 * if the this process is bigger than the biggest one
1342		 * remember it.
1343		 */
1344		if (size > bigsize) {
1345			if (bigproc != NULL)
1346				PROC_UNLOCK(bigproc);
1347			bigproc = p;
1348			bigsize = size;
1349		} else
1350			PROC_UNLOCK(p);
1351	}
1352	sx_sunlock(&allproc_lock);
1353	if (bigproc != NULL) {
1354		killproc(bigproc, "out of swap space");
1355		sched_nice(bigproc, PRIO_MIN);
1356		PROC_UNLOCK(bigproc);
1357		wakeup(&cnt.v_free_count);
1358	}
1359}
1360
1361/*
1362 * This routine tries to maintain the pseudo LRU active queue,
1363 * so that during long periods of time where there is no paging,
1364 * that some statistic accumulation still occurs.  This code
1365 * helps the situation where paging just starts to occur.
1366 */
1367static void
1368vm_pageout_page_stats()
1369{
1370	vm_object_t object;
1371	vm_page_t m,next;
1372	int pcount,tpcount;		/* Number of pages to check */
1373	static int fullintervalcount = 0;
1374	int page_shortage;
1375
1376	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1377	page_shortage =
1378	    (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1379	    (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1380
1381	if (page_shortage <= 0)
1382		return;
1383
1384	pcount = cnt.v_active_count;
1385	fullintervalcount += vm_pageout_stats_interval;
1386	if (fullintervalcount < vm_pageout_full_stats_interval) {
1387		tpcount = (int64_t)vm_pageout_stats_max * cnt.v_active_count /
1388		    cnt.v_page_count;
1389		if (pcount > tpcount)
1390			pcount = tpcount;
1391	} else {
1392		fullintervalcount = 0;
1393	}
1394
1395	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1396	while ((m != NULL) && (pcount-- > 0)) {
1397		int actcount;
1398
1399		KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE),
1400		    ("vm_pageout_page_stats: page %p isn't active", m));
1401
1402		next = TAILQ_NEXT(m, pageq);
1403		if ((m->flags & PG_MARKER) != 0) {
1404			m = next;
1405			continue;
1406		}
1407		vm_page_lock_assert(m, MA_NOTOWNED);
1408		if (!vm_pageout_page_lock(m, &next) ||
1409		    (object = m->object) == NULL) {
1410			vm_page_unlock(m);
1411			m = next;
1412			continue;
1413		}
1414		if (!VM_OBJECT_TRYLOCK(object) &&
1415		    !vm_pageout_fallback_object_lock(m, &next)) {
1416			VM_OBJECT_UNLOCK(object);
1417			vm_page_unlock(m);
1418			m = next;
1419			continue;
1420		}
1421
1422		/*
1423		 * Don't deactivate pages that are busy.
1424		 */
1425		if ((m->busy != 0) ||
1426		    (m->oflags & VPO_BUSY) ||
1427		    (m->hold_count != 0)) {
1428			vm_page_unlock(m);
1429			VM_OBJECT_UNLOCK(object);
1430			vm_page_requeue(m);
1431			m = next;
1432			continue;
1433		}
1434
1435		actcount = 0;
1436		if (m->flags & PG_REFERENCED) {
1437			vm_page_flag_clear(m, PG_REFERENCED);
1438			actcount += 1;
1439		}
1440
1441		actcount += pmap_ts_referenced(m);
1442		if (actcount) {
1443			m->act_count += ACT_ADVANCE + actcount;
1444			if (m->act_count > ACT_MAX)
1445				m->act_count = ACT_MAX;
1446			vm_page_requeue(m);
1447		} else {
1448			if (m->act_count == 0) {
1449				/*
1450				 * We turn off page access, so that we have
1451				 * more accurate RSS stats.  We don't do this
1452				 * in the normal page deactivation when the
1453				 * system is loaded VM wise, because the
1454				 * cost of the large number of page protect
1455				 * operations would be higher than the value
1456				 * of doing the operation.
1457				 */
1458				pmap_remove_all(m);
1459				vm_page_deactivate(m);
1460			} else {
1461				m->act_count -= min(m->act_count, ACT_DECLINE);
1462				vm_page_requeue(m);
1463			}
1464		}
1465		vm_page_unlock(m);
1466		VM_OBJECT_UNLOCK(object);
1467		m = next;
1468	}
1469}
1470
1471/*
1472 *	vm_pageout is the high level pageout daemon.
1473 */
1474static void
1475vm_pageout()
1476{
1477	int error, pass;
1478
1479	/*
1480	 * Initialize some paging parameters.
1481	 */
1482	cnt.v_interrupt_free_min = 2;
1483	if (cnt.v_page_count < 2000)
1484		vm_pageout_page_count = 8;
1485
1486	/*
1487	 * v_free_reserved needs to include enough for the largest
1488	 * swap pager structures plus enough for any pv_entry structs
1489	 * when paging.
1490	 */
1491	if (cnt.v_page_count > 1024)
1492		cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1493	else
1494		cnt.v_free_min = 4;
1495	cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1496	    cnt.v_interrupt_free_min;
1497	cnt.v_free_reserved = vm_pageout_page_count +
1498	    cnt.v_pageout_free_min + (cnt.v_page_count / 768);
1499	cnt.v_free_severe = cnt.v_free_min / 2;
1500	cnt.v_free_min += cnt.v_free_reserved;
1501	cnt.v_free_severe += cnt.v_free_reserved;
1502
1503	/*
1504	 * v_free_target and v_cache_min control pageout hysteresis.  Note
1505	 * that these are more a measure of the VM cache queue hysteresis
1506	 * then the VM free queue.  Specifically, v_free_target is the
1507	 * high water mark (free+cache pages).
1508	 *
1509	 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1510	 * low water mark, while v_free_min is the stop.  v_cache_min must
1511	 * be big enough to handle memory needs while the pageout daemon
1512	 * is signalled and run to free more pages.
1513	 */
1514	if (cnt.v_free_count > 6144)
1515		cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1516	else
1517		cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1518
1519	if (cnt.v_free_count > 2048) {
1520		cnt.v_cache_min = cnt.v_free_target;
1521		cnt.v_cache_max = 2 * cnt.v_cache_min;
1522		cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1523	} else {
1524		cnt.v_cache_min = 0;
1525		cnt.v_cache_max = 0;
1526		cnt.v_inactive_target = cnt.v_free_count / 4;
1527	}
1528	if (cnt.v_inactive_target > cnt.v_free_count / 3)
1529		cnt.v_inactive_target = cnt.v_free_count / 3;
1530
1531	/* XXX does not really belong here */
1532	if (vm_page_max_wired == 0)
1533		vm_page_max_wired = cnt.v_free_count / 3;
1534
1535	if (vm_pageout_stats_max == 0)
1536		vm_pageout_stats_max = cnt.v_free_target;
1537
1538	/*
1539	 * Set interval in seconds for stats scan.
1540	 */
1541	if (vm_pageout_stats_interval == 0)
1542		vm_pageout_stats_interval = 5;
1543	if (vm_pageout_full_stats_interval == 0)
1544		vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1545
1546	swap_pager_swap_init();
1547	pass = 0;
1548	/*
1549	 * The pageout daemon is never done, so loop forever.
1550	 */
1551	while (TRUE) {
1552		/*
1553		 * If we have enough free memory, wakeup waiters.  Do
1554		 * not clear vm_pages_needed until we reach our target,
1555		 * otherwise we may be woken up over and over again and
1556		 * waste a lot of cpu.
1557		 */
1558		mtx_lock(&vm_page_queue_free_mtx);
1559		if (vm_pages_needed && !vm_page_count_min()) {
1560			if (!vm_paging_needed())
1561				vm_pages_needed = 0;
1562			wakeup(&cnt.v_free_count);
1563		}
1564		if (vm_pages_needed) {
1565			/*
1566			 * Still not done, take a second pass without waiting
1567			 * (unlimited dirty cleaning), otherwise sleep a bit
1568			 * and try again.
1569			 */
1570			++pass;
1571			if (pass > 1)
1572				msleep(&vm_pages_needed,
1573				    &vm_page_queue_free_mtx, PVM, "psleep",
1574				    hz / 2);
1575		} else {
1576			/*
1577			 * Good enough, sleep & handle stats.  Prime the pass
1578			 * for the next run.
1579			 */
1580			if (pass > 1)
1581				pass = 1;
1582			else
1583				pass = 0;
1584			error = msleep(&vm_pages_needed,
1585			    &vm_page_queue_free_mtx, PVM, "psleep",
1586			    vm_pageout_stats_interval * hz);
1587			if (error && !vm_pages_needed) {
1588				mtx_unlock(&vm_page_queue_free_mtx);
1589				pass = 0;
1590				vm_page_lock_queues();
1591				vm_pageout_page_stats();
1592				vm_page_unlock_queues();
1593				continue;
1594			}
1595		}
1596		if (vm_pages_needed)
1597			cnt.v_pdwakeups++;
1598		mtx_unlock(&vm_page_queue_free_mtx);
1599		vm_pageout_scan(pass);
1600	}
1601}
1602
1603/*
1604 * Unless the free page queue lock is held by the caller, this function
1605 * should be regarded as advisory.  Specifically, the caller should
1606 * not msleep() on &cnt.v_free_count following this function unless
1607 * the free page queue lock is held until the msleep() is performed.
1608 */
1609void
1610pagedaemon_wakeup()
1611{
1612
1613	if (!vm_pages_needed && curthread->td_proc != pageproc) {
1614		vm_pages_needed = 1;
1615		wakeup(&vm_pages_needed);
1616	}
1617}
1618
1619#if !defined(NO_SWAPPING)
1620static void
1621vm_req_vmdaemon(int req)
1622{
1623	static int lastrun = 0;
1624
1625	mtx_lock(&vm_daemon_mtx);
1626	vm_pageout_req_swapout |= req;
1627	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1628		wakeup(&vm_daemon_needed);
1629		lastrun = ticks;
1630	}
1631	mtx_unlock(&vm_daemon_mtx);
1632}
1633
1634static void
1635vm_daemon()
1636{
1637	struct rlimit rsslim;
1638	struct proc *p;
1639	struct thread *td;
1640	struct vmspace *vm;
1641	int breakout, swapout_flags;
1642
1643	while (TRUE) {
1644		mtx_lock(&vm_daemon_mtx);
1645		msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0);
1646		swapout_flags = vm_pageout_req_swapout;
1647		vm_pageout_req_swapout = 0;
1648		mtx_unlock(&vm_daemon_mtx);
1649		if (swapout_flags)
1650			swapout_procs(swapout_flags);
1651
1652		/*
1653		 * scan the processes for exceeding their rlimits or if
1654		 * process is swapped out -- deactivate pages
1655		 */
1656		sx_slock(&allproc_lock);
1657		FOREACH_PROC_IN_SYSTEM(p) {
1658			vm_pindex_t limit, size;
1659
1660			/*
1661			 * if this is a system process or if we have already
1662			 * looked at this process, skip it.
1663			 */
1664			PROC_LOCK(p);
1665			if (p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
1666				PROC_UNLOCK(p);
1667				continue;
1668			}
1669			/*
1670			 * if the process is in a non-running type state,
1671			 * don't touch it.
1672			 */
1673			breakout = 0;
1674			FOREACH_THREAD_IN_PROC(p, td) {
1675				thread_lock(td);
1676				if (!TD_ON_RUNQ(td) &&
1677				    !TD_IS_RUNNING(td) &&
1678				    !TD_IS_SLEEPING(td)) {
1679					thread_unlock(td);
1680					breakout = 1;
1681					break;
1682				}
1683				thread_unlock(td);
1684			}
1685			if (breakout) {
1686				PROC_UNLOCK(p);
1687				continue;
1688			}
1689			/*
1690			 * get a limit
1691			 */
1692			lim_rlimit(p, RLIMIT_RSS, &rsslim);
1693			limit = OFF_TO_IDX(
1694			    qmin(rsslim.rlim_cur, rsslim.rlim_max));
1695
1696			/*
1697			 * let processes that are swapped out really be
1698			 * swapped out set the limit to nothing (will force a
1699			 * swap-out.)
1700			 */
1701			if ((p->p_flag & P_INMEM) == 0)
1702				limit = 0;	/* XXX */
1703			vm = vmspace_acquire_ref(p);
1704			PROC_UNLOCK(p);
1705			if (vm == NULL)
1706				continue;
1707
1708			size = vmspace_resident_count(vm);
1709			if (limit >= 0 && size >= limit) {
1710				vm_pageout_map_deactivate_pages(
1711				    &vm->vm_map, limit);
1712			}
1713			vmspace_free(vm);
1714		}
1715		sx_sunlock(&allproc_lock);
1716	}
1717}
1718#endif			/* !defined(NO_SWAPPING) */
1719